Search Images Maps Play YouTube News Gmail Drive More »
Sign in
Screen reader users: click this link for accessible mode. Accessible mode has the same essential features but works better with your reader.

Patents

  1. Advanced Patent Search
Publication numberUS7822885 B2
Publication typeGrant
Application numberUS 11/974,729
Publication dateOct 26, 2010
Filing dateOct 16, 2007
Priority dateOct 16, 2007
Fee statusPaid
Also published asUS8108571, US20090100200
Publication number11974729, 974729, US 7822885 B2, US 7822885B2, US-B2-7822885, US7822885 B2, US7822885B2
InventorsDaniel L. Bouvier
Original AssigneeApplied Micro Circuits Corporation
Export CitationBiBTeX, EndNote, RefMan
External Links: USPTO, USPTO Assignment, Espacenet
Channel-less multithreaded DMA controller
US 7822885 B2
Abstract
A channel-less system and method are provided for multithreaded communications with a direct memory access (DMA) controller. The method accepts a plurality of DMA command messages directed to a fixed port address. The DMA command messages are arranged in a first-in first-out (FIFO) queue, in the order in which they are received. The DMA command messages are supplied to a DMA controller from the FIFO queue, and in response to the DMA command message, data transfer operation are managed by the DMA controller. Following the completion of each data transfer operation, a transfer complete message indicating completion is sent. In one aspect, DMA command messages are arranged in a plurality of parallel FIFO queues, and CD sets are stored in a plurality of context memories, where each context memory is associated with a corresponding FIFO queue.
Images(6)
Previous page
Next page
Claims(14)
1. A method for multithreaded communications, the method comprising:
accepting a plurality of direct memory access (DMA) command messages directed to a fixed port address, each DMA command message including a completion notification field, with a notification address embedded in the completion notification field, and control descriptors (CDs), where the CDs are instruction sequences to be performed by a DMA completion engine;
arranging each DMA command message in one of a first fixed number plurality of parallel non-virtual first-in first-out (FIFO) queues, where each of the first fixed number plurality of non-virtual FIFO queues shares the fixed port address, and where each of the first fixed number plurality of non-virtual FIFO queues arranges DMA messages in an order in which they are received;
fetching a DMA command message from one of the first fixed number plurality of non-virtual FIFO queues;
storing the CDs from the DMA command message in one of a first plurality of context memories, where each context memory is associated with a corresponding one of the first fixed number plurality of non-virtual FIFO queues, and where each context memory has a priority ranking; and,
loading the DMA completion engine with CDs from one of the first plurality of context memories;
in response to performing the instruction sequences in the CDs, the DMA completion engine concurrently managing a plurality of data transfer operations, giving preferential treatment to CDs sourced from higher priority ranked context memories;
for each DMA command message, sending a transfer complete message to the notification address embedded in the DMA command message, indicating completion of an associated data transfer operation;
wherein accepting the plurality of DMA command messages includes accepting a first DMA command message and a second DMA command message;
wherein arranging the plurality of DMA command messages in the first fixed number plurality of non-virtual FIFO queues includes arranging the first DMA message in a first non-virtual FIFO queue and arranging the second DMA command message in the first non-virtual FIFO queue, subsequent to the first DMA command message;
wherein fetching the DMA command message from one of the first fixed number plurality of non-virtual FIFO queues includes fetching the second DMA command message from the first non-virtual FIFO queue, subsequent to fetching the first DMA command message from the first non-virtual FIFO queue; and,
wherein concurrently managing the plurality of data transfer operations in response to the CDs includes executing all CDs associated with the first DMA command message, prior to executing CDs associated with the second DMA command message.
2. The method of claim 1 wherein storing the CDs from the DMA command message includes storing CDs selected from a group consisting of direct CDs and pointer CDs.
3. The method of claim 1 wherein loading the DMA completion engine with CDs from one of the first plurality of context memories includes loading the DMA completion engine with CDs from context memories having a higher ranking using a higher message bandwidth.
4. The method of claim 1 wherein storing the CDs in one of the first plurality of context memories additionally includes:
storing CD state information; and,
storing the progress towards completion of a data transfer operation associated with the stored CD.
5. A system for multithreaded communications, the system comprising:
a first fixed number plurality of non-virtual first-in first-out (FIFO) queues sharing a fixed port address input for accepting direct memory access (DMA) command messages, each DMA command message having a completion notification field with a notification address embedded in the completion notification field and control descriptors (CDs), which are instructions sequences to be performed by a DMA completion engine, where each DMA command message is loaded into one of the first fixed number plurality of non-virtual FIFO queues, each of the first fixed number plurality of non-virtual FIFO queues arranging the DMA command messages in an order in which they are received and supplying the DMA command messages at a non-virtual FIFO queue output;
a CD fetch engine having an input for accepting the DMA command messages from the first fixed number plurality of non-virtual FIFO queues, the CD fetch engine recovering CDs from the DMA command messages and supplying the CDs at a CD fetch engine output; and,
a first plurality of context memories, each having an input to accept and store CDs, where each context memory is associated with a corresponding one of the first fixed number plurality of non-virtual FIFO queues, where each context memory has a priority ranking, and where each context memory supplies stored CDs at a corresponding context memory output; and,
the DMA completion engine having an interface to accept CDs from the context memories, giving preferential treatment to the CDs sourced from higher priority ranked context memories, the DMA completion engine having an input/output (IO) read/write interface for concurrently managing a plurality of data transfer operations, and a transfer complete interface for sending transfer complete messages to the notification addresses embedded in the completion notification fields, indicating the completion of the data transfer operations.
6. The system of claim 5 wherein the first plurality of context memories store CDs selected from a group consisting of direct CDs and pointer CDs.
7. The system of claim 5 wherein higher priority ranked context memories supply CDs to the completion engine at a higher message bandwidth than lower priority ranked context memories.
8. The system of claim 5 wherein the CD fetch engine supplies a second DMA command message, subsequent to a first DMA command message, to a first context memory; and,
wherein the DMA completion engine executes all CDs associated with the first DMA command message, prior to accepting CDs associated with the second DMA command message.
9. The system of claim 5 wherein the CD fetch engine supplies a first DMA command message to a first context memory, and a second DMA command message to a second context memory, having a higher priority ranking than the first context memory; and,
wherein the DMA completion engine executes CDs associated with the second DMA command message, prior to finishing the execution of CDs associated with the first DMA command message.
10. The system of claim 5 wherein the context memories store CD state information and progress of data transfer operations associated with the stored CDs.
11. A method for multithreaded communications, the method comprising:
accepting a plurality of direct memory access (DMA) command messages directed to a fixed port address, each DMA command message including a completion notification field, with a notification address embedded in the completion notification field, and control descriptors (CDs), where the CDs are instruction sequences to be performed by a DMA completion engine;
arranging each DMA command message in one of a first fixed number plurality of parallel non-virtual first-in first-out (FIFO) queues, where each of the first fixed number plurality of non-virtual FIFO queues shares the fixed port address, and where each of the first fixed number plurality of non-virtual FIFO queues arranges DMA messages in an order in which they are received;
fetching a DMA command message from one of the first fixed number plurality of non-virtual FIFO queues;
storing the CDs from the DMA command message in one of a first plurality of context memories, where each context memory is associated with a corresponding non-virtual FIFO queue, and where each context memory has a priority ranking; and,
loading the DMA completion engine with CDs from one of the first plurality of context memories;
in response to performing the instruction sequences in the CDs, the DMA completion engine concurrently managing a plurality of data transfer operations, giving preferential treatment to CDs sourced from higher priority ranked context memories;
for each DMA command message, sending a transfer complete message to the notification address embedded in the DMA command message, indicating completion of an associated data transfer operation;
wherein accepting the plurality of DMA command messages includes accepting a first DMA command message and a second DMA command message;
wherein arranging the plurality of DMA command messages in the first fixed number plurality of non-virtual FIFO queues includes arranging the first DMA command message in a first non-virtual FIFO queue and the second DMA command message in a second non-virtual FIFO queue;
wherein storing CDs from the DMA command message in one of the first plurality of context memories includes storing CDs from the first DMA command message in a first context memory, and storing CDs from the second DMA command message in a second context memory, having a higher priority ranking than the first context memory; and,
wherein concurrently managing the plurality of data transfer operations in response to the CDs includes executing CDs associated with the second DMA command message, prior to finishing the execution of CDs associated with the first DMA command message.
12. The method of claim 11 wherein storing the CDs from the DMA command message includes storing CDs selected from a group consisting of direct CDs and pointer CDs.
13. The method of claim 11 wherein loading the DMA completion engine with CDs from one of the first plurality of context memories includes loading the DMA completion engine with CDs from context memories having a higher ranking using a higher message bandwidth.
14. The method of claim 11 wherein storing the CDs in one of the first plurality of context memories additionally includes:
storing CD state information; and,
storing the progress towards completion of a data transfer operation associated with the stored CD.
Description
BACKGROUND OF THE INVENTION

1. Field of the Invention

This invention generally relates to Direct Access Memory (DMA) processing and, more particularly, to a channel-less multithreaded DMA controller.

2. Description of the Related Art

A general purpose programmable DMA controller is a software-managed programmable peripheral block charged with moving or copying data from one memory address to another memory address. The DMA controller provides a more efficient mechanism to perform large data block transfers, as compared to a conventional general purpose microprocessor. The employment of DMA controllers frees up the processor and software to perform other operations in parallel. Instruction sequences for the DMA, often referred to as control descriptors (CDs), are set up by software and usually include a source address, destination address, and other relevant transaction information. A DMA controller may perform other functions such as data manipulations or calculations.

Control descriptors are often assembled in groups called descriptor sequences or rings. Typically, the software control of a DMA controller is enabled through a device specific driver. The device driver is responsible for low level handshaking between upper layer software and the hardware. This device driver manages the descriptor rings, communicates with the DMA controller when work is pending, and communicates with upper layer software when work is complete.

It is possible that a DMA controller may be shared by many concurrent software threads running on one or more processors. Conventionally, a DMA controller maintains the logical concept of “channels”, whereby a channel provides the interface between a single software thread and the DMA controller. In other words, each software thread is associated with a channel. More concurrent driver threads require more DMA controller channels.

It has often been practice to provide multiple channels to address two concerns: thread sharing and quality of service. Oftentimes, multiple concurrent threads are used in systems for the parallel processing different aspects of data flow. Where multiple independent threads are deployed, sharing a common DMA controller can be cumbersome to manage. The software device driver in this case must not only provide the DMA controller with communications, but must also manage an arbitration scheme with upper layer software threads to determine which work gets done next. If this work is being carried out by multiple microprocessors, the overhead of coordination between threads is very complicated. The overhead coordination requires a certain level of software handshaking to determine which thread gets access to the controller at any particular time.

From a quality of service perspective, it is common to have higher and lower priority activities. For example, a software thread may queue a low priority transfer for a DMA controller. At some later time a different thread may be queued, which needs to run a higher priority task on the DMA controller. The ability to pre-empt low priority activities with higher priority tasks is a highly desired feature. Without such capability, a high priority operation must wait until a low priority operation is completed.

A multi-channel DMA controller addresses these issues where different software threads can be bound to specific channels and the underlying DMA controller hardware sorts out the access profile to memory based upon channel priorities. A disadvantage of the channel approach is that there are a limited number of hardware channels. If more logical threads exist than physical channels, then a software mechanism must once again be deployed to take care of the resource contention.

DMA controllers must also maintain a certain level of atomicity with respect to the execution of operations. This means that operational sequences must complete in the order programmed by software. This is typically accomplished using a run-to-completion model whereby a DMA channel completes all operations of a first CD, before moving onto the next CD. However, a brute-force run-to-completion methodology may prevent the data moving engine from performing unrelated operations (operations from different CD lists) in parallel, even if the engine is capable.

The communication between software and the DMA controller hardware is typically handled through a programmed input/output (IO) interface. That is, the software device driver programs control registers within the DMA channel, which causes the DMA to carry out the desired action. When the DMA controller is finished it communicates back to software, either through use of a hardware interrupt request, or through setting of a status bit that is polled by software. The software must wait until the current instruction sequence is complete before programming the next sequence. During the software/hardware handshake period the DMA channel is idle waiting for the next CD, thus resulting in dead time that could have been used for real work. To overcome this dead time, DMA controllers may deploy the concept of control descriptor sequences (CDS) and descriptor rings.

The descriptor ring provides a form of FIFO where the software adds new items in memory for the DMA channel at the tail of the ring, while the DMA controller processes CDs from the head of the ring. In this way, the software manages the tail pointers and the hardware (HW) manages the head pointer. Such schemes have the disadvantage of requiring software overhead to keep track of pointers.

It would be advantageous if the programming model limitations of having fixed DMA channels as main linkage between software and hardware could be eliminated, while preserving thread independence, intra-thread atomicity, and a quality of service capability.

SUMMARY OF THE INVENTION

To address the deficiencies of a channel-based DMA controller while maintaining all of its advantages, a channel-less multithreaded DMA controller is presented. The channel-less DMA controller makes use of hardware-managed command queues for which control descriptors are submitted. The software sends command messages to a DMA command queue, where they are in turn executed in the received order by the DMA controller. At the completion of each task a completion message is generated and sent back to software to indicate task completion. The queuing mechanism permits multiple commands to be pipelined to the DMA controller from many software thread sources, while still maintaining the run-to-completion atomic behavior. The software entry point to the command queue is a fixed port address or mailbox. Because the command queue is a fixed location, there are no software pointers that must be maintained and, therefore, multiple software threads can enqueue messages to the port without any awareness of each other. To maintain completion order and thread-level atomicity, the DMA controller pulls command messages from the command queue and processes them in. FIFO order run-to-completion.

Accordingly, a channel-less method is provided for multithreaded communications with a DMA controller. The method accepts a plurality of DMA command messages directed to a fixed port address. The DMA command messages are arranged in a FIFO queue, in the order in which they are received. The DMA command messages are supplied to a DMA controller from the FIFO queue, and in response to the DMA command message, data transfer operation are managed by the DMA controller. Following the completion of each data transfer operation, a transfer complete message indicating completion is sent.

Typically, the transfer complete message is sent to a software component. More explicitly, the DMA command messages may include a completion notification field with an embedded notification addresses. Then, the transfer complete message associated with a DMA command message is sent to the corresponding notification address.

In another aspect, DMA command messages are accessed from the FIFO queue. Sets of control descriptors (CDs) associated with (embedded in) the DMA command message are stored in a context memory, and a completion engine is loaded with sets of CDs from the context memory. In another aspect, DMA command messages are arranged in a plurality of parallel FIFO queues, and CD sets are stored in a plurality of context memories, where each context memory is associated with a corresponding FIFO queue.

Additional details of the above-described method and a channel-less system for multithreaded communications with a DMA controller are provided below.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a schematic block diagram of a channel-less system for multithreaded communications with a direct memory access (DMA) controller.

FIG. 2 is a schematic block diagram depicting a first variation in the system of FIG. 1.

FIG. 3 is a schematic block diagram depicting a second variation of the system of FIG. 1.

FIG. 4 is a schematic block diagram, depicting a third variation of the channel-less multithreaded DMA controller system.

FIG. 5 is a flowchart illustrating a channel-less method for multithreaded communications with a DMA controller.

DETAILED DESCRIPTION

Various embodiments are now described with reference to the drawings. In the following description, for purposes of explanation, numerous specific details are set forth in order to provide a thorough understanding of one or more aspects. It may be evident, however, that such embodiment(s) may be practiced without these specific details. In other instances, well-known structures and devices are shown in block diagram form in order to facilitate describing these embodiments.

As used in this application, the terms “processor”, “processing device”, “component,” “module,” “system,” and the like are intended to refer to a computer-related entity, either hardware, firmware, a combination of hardware and software, software, or software in execution. For example, a component may be, but is not limited to being, a process running on a processor, generated component, a processor, an object, an executable, a thread of execution, a program, and/or a computer. By way of illustration, both an application running on a computing device and the computing device can be a component. One or more components can reside within a process and/or thread of execution and a component may be localized on one computer and/or distributed between two or more computers. In addition, these components can execute from various computer readable media having various data structures stored thereon. The components may communicate by way of local and/or remote processes such as in accordance with a signal having one or more data packets (e.g., data from one component interacting with another component in a local system, distributed system, and/or across a network such as the Internet with other systems by way of the signal).

Various embodiments will be presented in terms of systems that may include a number of components, modules, and the like. It is to be understood and appreciated that the various systems may include additional components, modules, etc. and/or may not include all of the components, modules etc. discussed in connection with the figures. A combination of these approaches may also be used.

The various illustrative logical blocks, modules, and circuits that have been described may be implemented or performed with a general purpose processor, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any combination thereof designed to perform the functions described herein. A general-purpose processor may be a microprocessor, but in the alternative, the processor may be any conventional processor, controller, microcontroller, or state machine. A processor may also be implemented as a combination of computing devices, e.g., a combination of a DSP and a microprocessor, a plurality of microprocessors, one or more microprocessors in conjunction with a DSP core, or any other such configuration.

The methods or algorithms described in connection with the embodiments disclosed herein may be embodied directly in hardware, in a software module executed by a processor, or in a combination of the two. A software module may reside in RAM memory, flash memory, ROM memory, EPROM memory, EEPROM memory, registers, hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art. A storage medium may be coupled to the processor such that the processor can read information from, and write information to, the storage medium. In the alternative, the storage medium may be integral to the processor. The processor and the storage medium may reside in an ASIC. The ASIC may reside in the node, or elsewhere. In the alternative, the processor and the storage medium may reside as discrete components in the node, or elsewhere in an access network.

FIG. 1 is a schematic block diagram of a channel-less system for multithreaded communications with a direct memory access (DMA) controller. The system 100 comprises a first-in first-out (FIFO) queue 102 having a fixed port address input on line 104 for accepting DMA command messages. The FIFO queue 102 arranges the DMA command messages in the order in which they are received and supplies DMA command messages at an output on line 106. A DMA completion engine 108 has an interface on line 110 to accept control descriptors (CDs) embedded in the DMA commands, and an input/output (IO) read/write interface on line 112 for managing data transfer operations. A transfer complete interface on line 113 is used for sending transfer complete messages, indicating the completion of the data transfer operations. The communications transceived via the above-mentioned IO read/write interface would the communications performed by any conventional DMA controller in managing the transfer of data in memory (not shown).

In one aspect, the FIFO queue 102 accepts a first DMA command message 114 with a completion notification field 116 having an embedded first notification address and first set of CDs 118. The completion engine 108, upon completion of a first data transfer operation associated with a first set of CDs, sends a first transfer complete message to the first notification address via the interface on line 113.

The DMA command messages contains a field with an address to which the completion engine sends the completion message upon completion of the data management tasks associated with the DMA command message. This field may be referred to as a completion address field or completion notification field. There is no such field in conventional DMA controllers using the descriptor ring approach. Rather, notification information resides in the descriptor ring itself, and is associated with one, and only one software thread. The difference in notification between the system of FIG. 1 and a conventional descriptor ring DMA controller system is that the system of FIG. 1 permits multiple threads to share the command FIFO. In fact, since there is no need to track the notification information, multiple processors can now share the same command FIFO. The convention ring equivalent would require a single channel DMA controller with multiple threads adding work to a single ring. However, the only way to add multiple threads to a single ring is to use a software semaphore scheme to arbitrate between threads, so that different threads can add CDs to the ring one at a time. This software semaphore scheme is inefficient, as it complex and requires much overhead.

The system of FIG. 1 “abstracts” the hardware from the software, so that there is no hard and fast limitation between how many threads share the DMA engine. Note: the system of FIG. 1 is especially advantageous for multi-processor systems. With many cores, providing N number of channels to match a potential of N threads, a conventional DMA controller would require a significant amount of hardware. This is because each channel requires a dedicated set of context hardware to keep track of the ring or descriptor chain. With respect to FIG. 1, an equivalent conventional descriptor ring/DMA controller scheme would require N transfer complete paths, representing N channels. However, the system of FIG. 1 can process an infinite number of channels (limited only by the number of address bits for the completion message address) using a single FIFO component and a DMA command message with a completion notification field.

FIG. 2 is a schematic block diagram depicting a first variation in the system of FIG. 1. A CD fetch engine 200 has an input for accepting DMA command messages from the FIFO queue 102 on line 106. The CD fetch engine 200 recovers CD sets from DMA command messages and supplies the CD sets at an output on line 202. A context memory 204 has an input to accept the CD sets on line 202. The context memory stores the CD sets and supplies the stored CDs to the completion engine 108 at an output on line 110.

It should be understood that a CD set may include one or more CDs. A CD set may include direct CDs or pointer CDs. A direct CD embodies the command descriptor completely within the command message. A pointer CD contains a pointer to a memory location where a single command descriptor, or a linked list of command descriptors, is located. In this case, a single command message may actually represent a number of chained command descriptors.

FIG. 3 is a schematic block diagram depicting a second variation of the system of FIG. 1. In this aspect, a first plurality of parallel FIFO queues sharing the same fixed port address. The plurality of queues is represented as 102 a through 102 n, where n is not limited to any particular number. Also shown are a first plurality of context memories 204 a through 204 n, where each context memory is associated with a corresponding FIFO queue. Optionally, the first plurality of context memories 204 a-204 n may have a ranked order, with higher ranked context memories supplying CDs to the completion engine 108 at a higher message bandwidth than lower ranged context memories.

Because the DMA controller hardware is somewhat abstracted from the software, the controller can be scaled up or down without generally impacting the driver software. A larger more capable controller may have more parallel command queues while a smaller (lower cost) solution may have fewer. Since there is no hard channel binding between software and hardware, only the FIFO submit address needs to change. In a conventional hardware channel approach, each channel is hard-bound to the software driver. A reduction in the number of channels requires that the software be restructured.

In one aspect, FIFO queue 102 a supplies a second DMA command message 302, subsequent to a first DMA command message 300, to context memory 204 a. In accordance with the principle of atomicity, the completion engine 108 executes all CDs associated with the first DMA command message, prior to accepting CDs associated with the second DMA command message, from the first context memory 204 a.

In another aspect, FIFO queue 102 b supplies a third DMA command message 304 to context memory 204 b, and FIFO queue 102 n supplies a fourth DMA command message 306 to context memory 204 n. In circumstances where atomicity is not required, the completion engine 108 executes CDs supplied from context memory 204 n, prior to finishing the execution of CDs supplied from context memory 204 b. Alternately, the completion engine 108 may execute CDs supplied from context memory 204 b, prior to finishing the execution of CDs supplied from context memory 204 n.

In another aspect of the system, context memories 204 a-204 nmay store CD state information and the progress towards competition of the stored CDs, as explained in more detail below. Although the system is depicted as a network of connected hardware blocks, it should be understood that components of the system may be enabled using processor instructions stored in memory and executed by a processor device, or enabled as a logic-coded state machine. Although the system has been depicted as a combination of separate components, in some aspects some or all of the separate components may be embedded in a single chip, or embedded with components (not shown) from other systems.

Functional Description

The channel-less DMA controller makes use of hardware-managed command queues for which control descriptors are submitted. The software sends command messages to a DMA command FIFO queue, where they are in turn executed in the received order by the DMA controller. At the completion of each task a completion message is generated and sent back to software to indicate task completion. The FIFO queue permits multiple commands to be pipelined to the DMA controller from many software thread sources, while still maintaining the run-to-completion atomic behavior. The software entry point to the command queue is a fixed port address or mailbox. Because the command queue is a fixed location, there are no software pointers that must be maintained and, therefore, multiple software threads can enqueue messages to the port without any awareness of each other. To maintain completion order and thread level atomicity, the DMA controller pulls command messages from the command queue and processes them in FIFO order run-to-completion.

As a variation, multiple command queues may be employed to provide quality of service operations. Queues, or more particularly, the context memories being serviced by the queues, may be assigned different priority ranks. Commands that are posted to a higher priority queue are given preferential service by the DMA controller. This preferential service is typically done in the form of more transactional bandwidth allocation. To accomplish pre-emption, the DMA controller manages multiple thread contexts, thus keeping track of where each thread is at any point in time.

The DMA controller works in two operational modes: direct and chaining. In the direct mode, a software thread directly programs the DMA controller with the necessary transfer attributes, such as source and destination address, followed by a command to start the transfer. In the chaining or pointer mode, the software pre-loads into memory a list of transfers in the form of a command descriptor chain. This is followed by a command (pointer) to the DMA controller to begin execution with an address pointer to the start of the memory located chain.

One advantage of the channel-less DMA approach is the elimination of the conventional fixed-per-thread software-to-hardware binding for control and management of the DMA engine. Eliminated are the hardware constraints of not having enough “channels” to satisfy the number of unique software threads that would like to use the DMA engine. Also eliminated is the need for a centralized DMA software management thread through which other threads must communicate if not enough channels exist. Other advantages include the elimination of software overhead in managing a tail pointer or other mechanism for command enqueue to DMA controller. Software overhead is eliminated in managing quality of service (QoS), instead providing a hardware-managed QoS scheme while still maintaining per thread atomicity.

While the channel-less DMA approach may require additional hardware to contain and manage command queues for the DMA engine, this additional hardware is a small price to pay for the net gains in performance and flexibility to software. That is, the use of multiple command queues permits a quality of service scheme that can be implementing independent of software management, using the concept of higher priority and lower priority queues.

The DMA controller executes commands encoded by software into a command descriptor (CD). The CD describes the source and destination information for the transaction. In addition, the CD contains a command opcode and any other necessary attribute information to complete the transaction. CDs may be directly written to the DMA controller one at a time in the form of a command message. This is known as “direct mode”. Another option is to place the CD in memory and to send the DMA controller a message with a pointer address to the CD. This is known as “pointer mode”. In pointer mode, multiple CDs can be linked together to form a linked-list in memory for which the DMA controller will run in order to completion. This can be very useful for completing more complex linked operations.

The common usage model is for software to have pre-assembled message and CD templates in memory. The templates are then copied and modified with any unique parameters such as address pointers and then enqueued to one of the DMA message queues.

FIG. 4 is a schematic block diagram, depicting a third variation of the channel-less multithreaded DMA controller system. The DMA module maintains any number of message queues. Eight such queues (102) are shown in the figure. A software thread can generate command messages to any one of the queues. Commands within each queue are serviced in FIFO order and run-to-completion. The FIFO ordering forms atomicity between each successive command message within the queue, permitting the software to enqueue multiple commands without concern for completion order. There is no completion order enforcement for commands enqueued to different message queues. If completion order is required, the software must either enqueue commands to the same queue or must wait until a completion message is received before enqueuing the next message to a different queue.

While the DMA controller is executing a CD it maintains information regarding the operation. This information is known as “CD context”. A CD context may include the current byte count, address pointer, and next CD address pointer for linked-list operations.

For each of the eight message queues shown, the DMA controller maintains a pending CD context 204. This is the context information related to DMA operation that is currently in progress for that queue. In addition, the DMA module holds a prefetch CD context, which is the next CD operation that will be executed for the particular queue once the current operation is completed. Prefetching the CD context helps to pipeline operations to the DMA controller to minimize idle time.

The DMA controller can concurrently service commands from each command queue for up to eight simultaneous CDs in this example. This concurrency permits the servicing of higher priority operations without the need for waiting for lower priority operations to complete. The DMA controller initiates transactions for any of the current pending CD contexts. A weighted round robin scheme may be used, for example, to service the pending CD contexts, with each queue making at least one bus transaction before moving to the next queue. A transaction is defined as a single IO read bus transaction. Operations associated with higher priority queues are given more transactions on the IO read bus, thus completing sooner than operations associated with lower priority queues. A weighted round robin scheme guarantees that lower priority queues do not starve.

A processor generates command messages to one of the eight DMA message queues in this example. The DMA message queues are pre-programmed with a priority setting. The DMA Message Manager 400 extracts messages from the eight queues one at a time. If messages are pending in more than one queue a fairness scheme may be used to determine the order of queue extraction. The DMA message manager helps facilitate faster setup and servicing of messages from higher priority queues. Message entries (DMA command messages) within a queue are serviced by the DMA controller in FIFO order by the DMA message manager.

The extracted DMA message will either be a direct DMA command or a pointer DMA command. In the case of the direct DMA command, the message contains a CD. The CD includes the command opcode, source and destination information, and any other necessary information to complete the transaction. The information is directly loaded into the prefetch CD context store associated with the message queue. For pointer commands, the CD Fetch block 200 requests the CD context from the memory location pointed to by the address pointer in the message. The CD is fetched using the 10 AXI read port 112 and then loaded into the prefetch CD context store 402. Upon completion of the current pending CD context the prefetched context is loaded into pending context store 204 and the controller can begin servicing the CD. If the CD is a linked list, the CD Fetch block 200 fetches the next CD into the prefetch CD context store. This is done until the last link in the chain is reached, at which point the CD Fetch block and extract another message from the associated queue.

There are eight pending and eight active CD context stores in this example, one for each message queue. For direct DMA commands, when the prefetched CD context moves to the pending CD context, the CD Fetch block extracts the next message from the associated queue.

Upon reception of data for a fetch, the data enters into the completion engine 108. In the simplest form, the completion engine merely writes the data to the destination address through the IO write bus 112 and increments the destination address for the next transaction. The completion engine might also perform some data manipulations as supported and requested.

The DMA controller can be thought of as a fine-grain multithreaded processor where each pending CD context is considered a thread. The DMA controller executes only one fetch and one completion transaction at a time. A transaction is defined as one IO bus operation. If multiple pending contexts are present, then the fetch engine services them one transaction at a time, based on the weighted round robin scheme for example. Pending contexts from higher priority queues have more transaction issue slots than lower priority queues. Since data may return from fetches out of order, the completion engine operates on the pending contexts in the order that data is received.

The system of FIG. 4 eliminates need for the hardware-based binding of channels to threads. The system eliminates software overhead of managing a tail pointer or other mechanism for command enqueuing to DMA controller. The system provides a hardware managed QoS scheme while still maintaining per thread atomicity, thus eliminating software overhead to manage such.

A performance advantage of this system over a conventional channel-based controller can be envisioned as a logical pipeline. In a channel-based approach the software driver must first submit work to the channel in the form of a command descriptor. Next, the channel carries out the command descriptor followed by notifying the software driver that it is done (either through interrupt or software polling). This handshaking can result in large bubbles of DMA idle time in-between when the engine completes the operation and the next command descriptor is loaded. To avoid this problem descriptor rings were created, whereby software adds new command descriptors to the tail of the ring while the engine processes them on the head of the ring. The problem with the descriptor ring approach is that software must constantly check the next entry in the ring to see if the tail has caught the ring. This handshaking can also be expensive in performance. In the channel-less system the hard link between software and channel is not required, such that the software can submit any number of command messages to the engine. The software is notified at some later time when the operations are complete.

FIG. 5 is a flowchart illustrating a channel-less method for multithreaded communications with a DMA controller. Although the method is depicted as a sequence of numbered steps for clarity, the numbering does not necessarily dictate the order of the steps. It should be understood that some of these steps may be skipped, performed in parallel, or performed without the requirement of maintaining a strict order of sequence. The method starts at Step 500.

Step 502 accepts a plurality of DMA command messages directed to a fixed port address. Step 504 arranges the DMA command messages in a FIFO queue, in the order in which they are received. Step 506 supplies a first DMA command message to a DMA controller from the FIFO queue. In response to the first DMA command message, Step 508 manages a first data transfer operation. Step 510 sends a first transfer complete message indicating completion of the first data transfer operation.

In one aspect, supplying the first DMA command message to the DMA controller in Step 506 includes supplying a first DMA command message with a completion notification field having an embedded first notification address. Then, sending the first transfer complete message in Step 510 includes sending the first transfer complete message to the first notification address.

In another aspect, supplying the first DMA command message to the DMA controller (Step 506) includes substeps. Step 506 a fetches the first DMA command message from the FIFO queue. Step 506 b stores a first set of control descriptors (CDs) associated with the first DMA command message in a context memory. The CD set may include direct CDs and pointer CDs. Further, Step 506 b may store CD state information and the progress towards competition of the stored CDs. Step 560 c loads a completion engine with the first set of CDs from the context memory.

In one variation, arranging the DMA command messages in the FIFO queue in Step 504 includes arranging the DMA command messages in a first plurality of parallel FIFO queues. Then, storing the first set of CDs associated with the first DMA command message in a context memory (Step 506 b) includes storing a plurality of CD sets in a first plurality of context memories, where each context memory is associated with a corresponding FIFO queue. Further, the first plurality of context memories may be ranked, and loading the completion engine from the context memory (Step 506 c) may include providing a higher message bandwidth to context memories having a higher ranking.

In another variation, Step 506 supplies a second DMA command message, subsequent to the first DMA command message, to a first FIFO queue. Then, managing the first data transfer operation in response to the first DMA command message (Step 508) includes executing all CDs associated with the first DMA command message, prior to executing CDs associated with the second DMA command message.

In another aspect, Step 506 supplies the first DMA command message from a second FIFO queue, and a third DMA command message is supplied from a third FIFO queue. Then, Step 508 executes CDs associated with the third DMA command message, prior to finishing the execution of CDs associated with the first DMA command message.

A channel-less multithreaded DMA controller has been presented. Examples of particular components and process steps have been given to illustrate the invention. However, the invention is not limited to merely these examples. Other variations and embodiments of the invention will occur to those skilled in the art.

Patent Citations
Cited PatentFiling datePublication dateApplicantTitle
US5379399 *May 8, 1992Jan 3, 1995International Business Machines CorporationFIFO memory controller for a digital video communications channel having a detector, comparator, and threshold select logic circuit
US5444853 *Mar 31, 1992Aug 22, 1995Seiko Epson CorporationSystem and method for transferring data between a plurality of virtual FIFO's and a peripheral via a hardware FIFO and selectively updating control information associated with the virtual FIFO's
US5659749 *May 8, 1995Aug 19, 1997National Instruments CorporationSystem and method for performing efficient hardware context switching in an instrumentation system
US6128674 *Aug 8, 1997Oct 3, 2000International Business Machines CorporationMethod of minimizing host CPU utilization in driving an adapter by residing in system memory a command/status block a soft interrupt block and a status block queue
US6178462 *Nov 24, 1997Jan 23, 2001International Business Machines CorporationProtocol for using a PCI interface for connecting networks
US6480942 *May 27, 1999Nov 12, 2002Sony CorporationSynchronized FIFO memory circuit
US6549960 *Nov 16, 2000Apr 15, 2003International Business Machines CorporationArchitecture and apparatus for implementing 100 MBPS and GBPS ethernet address
US6789056 *Dec 31, 2002Sep 7, 2004Intel CorporationDSP interface for packet processing
US7099328 *Jul 31, 2001Aug 29, 2006Mindspeed Technologies, Inc.Method for automatic resource reservation and communication that facilitates using multiple processing events for a single processing task
US7120708 *Jun 30, 2003Oct 10, 2006Intel CorporationReaddressable virtual DMA control and status registers
US7424556 *Oct 7, 2005Sep 9, 2008Adaptec, Inc.Method and system for sharing a receive buffer RAM with a single DMA engine among multiple context engines
US7565462 *Nov 27, 2007Jul 21, 2009Broadcom CorporationMemory access engine having multi-level command structure
US7631128 *Jun 28, 2007Dec 8, 2009Emc CorporationProtocol controller for a data storage system
US20020075884 *Dec 12, 2001Jun 20, 2002Jeroen WellenDistributed scheduler for packet switches and passive optical networks
US20050080943 *Oct 9, 2003Apr 14, 2005International Business Machines CorporationMethod and apparatus for efficient sharing of DMA resource
US20070174513 *Jan 23, 2006Jul 26, 2007Arm LimitedBuffering data during a data transfer
US20090006546 *Jun 26, 2007Jan 1, 2009International Business Machines CorporationMultiple node remote messaging
Referenced by
Citing PatentFiling datePublication dateApplicantTitle
US8230136 *Nov 19, 2010Jul 24, 2012International Business Machines CorporationMultithreaded programmable direct memory access engine
US20110242119 *Apr 1, 2011Oct 6, 2011Bolz Jeffrey AGPU Work Creation and Stateless Graphics in OPENGL
Classifications
U.S. Classification710/22, 710/39, 710/19, 710/35, 710/42, 710/308, 710/34, 710/310, 710/305, 710/33, 710/36, 712/228, 710/52, 710/24, 710/40, 712/229, 710/23, 710/5, 718/108
International ClassificationG06F13/28, G06F9/46, G06F13/00, G06F13/14, G06F9/44, G06F9/00
Cooperative ClassificationG06F13/28
European ClassificationG06F13/28
Legal Events
DateCodeEventDescription
Apr 28, 2014FPAYFee payment
Year of fee payment: 4
Oct 16, 2007ASAssignment
Owner name: APPLIED MICRO CIRCUITS CORPORATION, CALIFORNIA
Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNOR:BOUVIER, DANIEL L.;REEL/FRAME:020029/0657
Effective date: 20071012